Feature #1505
Fires on “ is” (not labeled yet: read the examples)
- Fires on
- 0.78% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 91%
- story 4%
- agent 3%
- worksheet 2%
- python 0%
Makes the model say
. realized instead! - beakfact<|assistant_end|>
Makes it not say
HiOne But talles0Cane
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.18<|tool_call_end|><|assistant_end|><|tool_start|>320<|tool_end|><|assistant_start|>The answer is 320
-
chat 4.17619<|tool_call_end|><|assistant_end|><|tool_start|>1280<|tool_end|><|assistant_start|>The answer is 128
-
chat 4.07+ 93<|tool_call_end|><|assistant_end|><|tool_start|>175<|tool_end|><|assistant_start|>The answer is 175
-
chat 4.06+ 35<|tool_call_end|><|assistant_end|><|tool_start|>432<|tool_end|><|assistant_start|>The answer is 432
-
chat 4.06+ 34<|tool_call_end|><|assistant_end|><|tool_start|>756<|tool_end|><|assistant_start|>The answer is 756
-
chat 4.05+ 6<|tool_call_end|><|assistant_end|><|tool_start|>18<|tool_end|><|assistant_start|>The answer is 18.
-
chat 4.03+0=1, 131<|think_end|>The answer is 131
-
chat 4.010+7=7, 766<|think_end|>The answer is 766
-
chat 3.9746<|tool_call_end|><|assistant_end|><|tool_start|>46<|tool_end|><|assistant_start|>The answer is 46.
-
chat 3.96<|tool_call_end|><|assistant_end|><|tool_start|>50<|tool_end|><|assistant_start|>The answer is 50.
-
chat 3.941+0=1, 140<|think_end|>The answer is 140
-
chat 3.92=6, 65<|think_end|>The answer is 65.